The impact of artificial intelligence on numerous industries is breathtaking, and climate science is not left out in this transformation. As environmental scientists adopt powerful AI models to predict variations in climate and weather conditions, research indicates that bigger isn’t always better – particularly in climatic predictions.
A team of researchers from MIT revealed that smaller, physics-based models may surpass deep-learning models in predicting certain aspects of the climate. The team found that these traditional models were more precise while figuring out regional surface temperatures, contrary to the expectation of complicated AI models delivering more accurate results.
Noelle Selin, a professor at MIT’s Institute for Data, Systems, and Society, emphasized the need to create models relevant to policymakers. According to her, while AI may tempt scientists, it’s crucial to remember the problem’s essence before jumping on the AI bandwagon.
Interestingly, the research team stumbled upon biased evaluations of AI models. Natural fluctuations in climate data can distort the outcomes, overhyping the accuracy of these models. This prompted the researchers to develop a more robust method of evaluation. While they discovered that linear pattern scaling (LPS) performed better than complex models in predicting temperature ranges, deep-learning showed more promise in local rainfall predictions.
Climate emulators, handy tools in policy creation, simulate human impact on future climate conditions. They serve as quicker alternatives to full-blown climate models, but their accuracy matters the most. Comparing LPS and deep-learning using a widely accepted dataset, MIT researchers found that LPS outperformed deep learning on nearly all parameters, including precipitation and temperature.
Lead author, Björn Lütjens, a research scientist at IBM Research, further stressed that although large AI methods excite scientists, simpler solutions must be implemented first. He noted that a few results, like precipitation data, were at odds with initial expectations. Where it was assumed that deep learning models would fare better because of the non-linear pattern of rainfall, it struggled with long-term climate changes, making LPS the favored model.
To provide a more accurate picture, researchers formulated a new evaluative framework that accounted for natural climate variability. In this context, deep-learning marginally outperformed LPS in predicting local rainfall, while LPS remained the go-to model for temperature predictions. The researchers subsequently factored LPS into a climate emulation platform to improve localized temperature predictions across different emissions scenarios.
The goal of this research, as noted by co-author Raffaele Ferrari, isn’t to declare one method as superior, but to stress the value of appropriate tools for specific problems. The researchers hope that through their work, more improved benchmarking techniques would emerge, enabling researchers to discern the most suitable models for various climate prediction tasks.
Lütjens is optimistic that with an enhanced climate emulation benchmark, more complex machine-learning models can tackle problems currently difficult to solve, such as the impacts of aerosols or estimating extreme precipitation. This remarkable research is part of MIT’s Climate Grand Challenges initiative and was partially sponsored by Schmidt Sciences, LLC.
For more details, you can check out the full study on the Journal of Advances in Modeling Earth Systems or access the original article here.
This website uses cookies.